If I misjudge at first and hand a question that Chat could've answered over to Cowork instead, what's the actual cost?
The most direct cost is time — Cowork needs some startup and processing time to run a task, so if a question could've been answered by Chat in five seconds, sending it to Cowork instead means waiting longer for the same answer. That's exactly the "opposite mistake" the article calls out by name. Beyond time, this kind of misjudgment usually doesn't cause any other real harm — you won't end up with wrong data or a failed task just because you picked the wrong tool; it's purely a matter of wasted efficiency.
That's also why Anthropic doesn't treat "think it through before choosing" as some high-stakes decision — it's more a matter of building familiarity. The judgment criteria themselves aren't complicated, and intuition builds naturally with use. The cost of getting it wrong the first few times is low, so there's no need to over-deliberate out of fear of choosing wrong — actually trying it a few times and feeling out the difference builds accurate intuition faster than repeatedly parsing the criteria in the abstract.
For the criterion "you already know what good looks like," if I'm genuinely unfamiliar with a piece of work, does that mean I shouldn't use Cowork at all?
Not at all — it just means the risk assessment shifts in that situation. The point of this particular criterion is letting you quickly judge whether the output is right. If you're genuinely unfamiliar with the work itself, and don't even know what "good" looks like yourself, then even if Cowork produces a result, you won't be able to tell in 15 seconds whether it's correct. The risk here isn't about whether Cowork executed well — it's that you're not equipped to sign off on it.
The more practical approach is splitting this into two stages: if you're genuinely unfamiliar with the work itself, first use Chat to ask "what should the output for this look like, and what's the standard for judging good versus bad." Once you have a basic grasp of that standard, hand the actual execution to Cowork. In other words, "you already know what good looks like" isn't a hard gate that blocks you from using Cowork if you don't meet it — it's a reminder that if you don't have that knowledge yet, fill that gap first before delegating execution.
Anthropic recommends adding a line like "repeat my request back and ask clarifying questions" to a prompt — how exactly should that be phrased to actually get Claude to ask before starting?
The exact example Anthropic's article gives is: "Before we begin, repeat my ask back to me so we're aligned, then ask me as many clarifying questions as you have." The key is that this does two things at once — first requesting a repeat-back (confirming Claude's understanding matches what you actually meant), then requesting proactive questions (letting Claude surface information it needs that you haven't provided yet) — rather than doing just one of the two. If you only ask for a repeat-back, Claude might faithfully restate a request that was itself vague to begin with. If you only ask it to question you, you might skip the more basic calibration step of confirming both sides are on the same page first.
In practice, this line can be added verbatim to the end of any prompt where you're delegating a task, turning it into a standard opening move. Spending thirty seconds answering the few questions Claude asks back usually surfaces details you hadn't even realized needed clarifying (say, whether "last month" here means the calendar month or the past 30 days) — which saves far more time than discovering a mismatch after the task has already run and having to redo it.
If I'm not sure which of my regular tasks would actually suit Cowork, how do I actually phrase Anthropic's suggestion to "just ask Claude"?
What Anthropic's article describes is leveraging the fact that Claude has memory and can search past conversations, and directly asking it to analyze the kinds of work you've asked it to handle before. A concrete way to phrase it might be something like: "Look back through our past conversations and identify the few types of work I most often ask you to help with, and for each one, judge whether it's better suited to a one-off Chat exchange or to becoming a recurring task in Cowork." This shifts the basis for the judgment from your own subjective recollection of "I feel like I do this a lot" to something grounded in more objective data — your actual conversation history.
This approach works especially well for someone who's already been using Claude for a while and has built up a decent amount of conversation history. If you're just getting started, your past conversation history is thin to begin with, so this method has limited reference value. In that case, the more practical approach is going straight through the five-point checklist and comparing it against a few things you're currently working on, rather than relying on retrospective usage patterns.
The place beginners most often get stuck usually isn't "how do I operate Cowork" — it's "should this particular thing be done in Cowork, or should I just open a regular chat window and ask." Anthropic's own growth marketing lead calls out, in an official blog post, the two opposite mistakes most people make: one is reaching for chat for everything and never actually feeling the difference Cowork can make; the other is the reverse — handing Cowork a simple question and waiting around for it to finish, when the same question dropped into a chat window would've had an answer in five seconds. This piece breaks down Anthropic's own decision framework so you can avoid both extremes.
Anthropic's blog post sums up the division of labor in one line: Chat is you bringing what you have to Claude — upload a file, paste some text, describe what's going on, and get an answer back. It's for getting answers, brainstorming, or thinking out loud. Cowork flips that around — you bring Claude to your work instead: point it at a folder on your computer, connect the apps you already use, and tell it what outcome you want. You describe an outcome, step away, and come back to finished work. Claude Code is built for developers shipping software — if your work lives in code, start there. Most knowledge work (email, decks, spreadsheets, docs, meetings, "can you pull together a summary of this") falls under the first two categories.
The article condenses "which one, when" into a single rule: if what you want fits in a few exchanges — a question, an explanation, a brainstorm, or a gut check — use Chat. If what you need is a deliverable, like a file someone will open, a deck someone will present, or a spreadsheet that needs sorting, use Cowork. Anything multi-step, touching more than one file or app, or something you'd describe as "a task" rather than "a question" falls under Cowork's territory — in that case, you're delegating work to Claude, not chatting with it.
Anthropic's blog post includes a short table of sample questions versus tasks, with a few examples that draw the line clearly: "What should I cover in our business review meeting?" — Chat. "Read the last three months of meeting notes in this Google Drive folder and build me a QBR deck using our template" — Cowork. "How do I VLOOKUP something?" — Chat. "Go through my spreadsheets and change all the VLOOKUP to INDEX MATCH" — Cowork. "Suggest a better title tag and meta description for this page" — Chat. "Use the new title tags and meta descriptions for these 30 pages from this sheet and update them using the CMS connector" — Cowork. What these pairs have in common: the Cowork side always involves actually reading, modifying, or producing something concrete, while the Chat side stays at the level of discussion or suggestion.
If gut feel still isn't enough to decide, Anthropic offers a five-point checklist — you don't need to hit all five; hitting a few makes a task a good candidate for Cowork. First, more than one thing goes in — multiple files, a whole folder, or a file plus some connectors; if there's only one input, Chat usually handles it fine. Second, a file comes out — you need a deliverable you can attach, present, share, or repurpose: a doc, a deck, a spreadsheet, or a CSV. Third, you'll do it again — one-offs are fine, but recurring tasks are where Cowork's value really shows, to the point you can schedule them to run before you're even at your desk. Fourth, you already know what good looks like — you're familiar enough with the shape of the output that you can tell in about 15 seconds whether it's right, wrong, or 70% there. Fifth, the middle is the boring part — the actual thinking happens at the start (deciding what you want) and the end (judging if it's right); everything in between — extracting, compiling, reconciling, reformatting — is what you actually want to hand off.
The blog post lays out a concrete ten-minute onboarding flow: open the Claude Desktop app and switch to the Cowork tab; give Claude something to work with — drop in a few files, point it at a folder on your computer, or connect an app you frequently use (Slack, Gmail, Notion, a CRM); clearly state the outcome you want, describing the deliverable and any necessary context; start with a real task you already know well, so you can immediately see where it's strong, where it needs more context from you, and you already have your own yardstick for what "good" looks like. The last one is what Anthropic specifically calls out as the single most useful habit to build: add a line to your prompt asking Claude to repeat your request back to confirm Alignment before starting, then ask as many clarifying questions as it has. This surfaces details you didn't think to specify — which time period you actually mean, what "good" means in this context, or an Edge Case you know about that Claude doesn't. Spending thirty seconds answering a few questions upfront costs far less than discovering those same gaps afterward and spending time and tokens fixing them.
Next time you're unsure how to classify a task, ask yourself: can this be said in a few sentences? Or am I actually describing a sequence of steps, touching more than one file or app, that'll produce something concrete once it's done? If the answer leans toward the latter, that's almost certainly Cowork's territory. If you're still not sure, Anthropic's own documentation has a genuinely practical suggestion: just ask Claude. Claude has memory and can search your past conversations, so you can ask it to help identify which tasks you do most often and which ones are worth trying in Cowork — rather than guessing on your own, let it find the answer from your actual usage patterns.